Presentation
Codesign Supercomputer Fugaku and Target Applications through Performance Optimization
SessionResearch Poster Session
Event Type
Research Poster
Pre-Recorded
TimeTuesday, June 23rd3:15pm - 3:20pm
LocationAnalog 1
DescriptionIn the development of Fugaku - the flagship supercomputer of Japan -
codesign with target applications plays an essential role.
The target applications, representing the nine social and scientific priority issues
chosen by the ministry, are used throughout the design and implementation phases
of Fugaku. They cover different numerical schemes, space and time discretization,
grid structure and data types. They have different workload characteristics
and exhibit the bottlenecks in the system's different components.
Addressing and relaxing these performance bottlenecks should improve
the computational performance of the applications for not only the priority issues
but also the wide range of high performance computing demand on Fugaku.
Fugaku performance prediction is conducted using various performance tools
available at each phase of development. Such tools include
the performance prediction tools based on FX100 hardware performance counters,
the software simulators at cycle level precision, hardware emulator and
Fugaku prototype test vehicle.
Our consensus is that formulating the general sensitivity in the systems design
parameter space is not feasible, and that experimenting the best path forward approach
is only realistic.
The codesign is effectively the repeated mutual optimization between the application
and the system in terms of performance, with power consumption and economy restriction.
From systems view, the codesign in the early stage contributes to the design of
system architecture such as the processor and memory specification and configuration,
and in the later stage, it contributes to the design of the system software such as
compilers and libraries.
codesign with target applications plays an essential role.
The target applications, representing the nine social and scientific priority issues
chosen by the ministry, are used throughout the design and implementation phases
of Fugaku. They cover different numerical schemes, space and time discretization,
grid structure and data types. They have different workload characteristics
and exhibit the bottlenecks in the system's different components.
Addressing and relaxing these performance bottlenecks should improve
the computational performance of the applications for not only the priority issues
but also the wide range of high performance computing demand on Fugaku.
Fugaku performance prediction is conducted using various performance tools
available at each phase of development. Such tools include
the performance prediction tools based on FX100 hardware performance counters,
the software simulators at cycle level precision, hardware emulator and
Fugaku prototype test vehicle.
Our consensus is that formulating the general sensitivity in the systems design
parameter space is not feasible, and that experimenting the best path forward approach
is only realistic.
The codesign is effectively the repeated mutual optimization between the application
and the system in terms of performance, with power consumption and economy restriction.
From systems view, the codesign in the early stage contributes to the design of
system architecture such as the processor and memory specification and configuration,
and in the later stage, it contributes to the design of the system software such as
compilers and libraries.
Poster PDF